a Dify tool for storing and retrieving long-term-memory, using Dify built-in Knowledge dataset for storing memories, each user has a standalone long-term-memory space. Make chatbot can persistently remember customized information per user.

workflow view:

Here is the major logic of this workflow:

1. before each conversation run a "HTTP Request" node named "Action:retrieve_memory_of_user" to request Dify Knowledge API to query user related document segments with filter parameters as keyword=$current_user_id

2. run a "Template" node as "Agent prompt"(agent prompt content) for preparing agent system prompt, injected customized $memory_template

3. run a "Parameter Extractor" node to extract $action(value will be one of reply/add_user_memory/update_user_memory ) and $longterm_memory_content parameters

4. if the $action is not "reply", run a "HTTP Request" node to request Dify Knowledge API to create or update the document segment with $longterm_memory_content parameter

5. if the $action is "reply", return the retrieved segments as a output variable $longterm_memory_space, so that it can be injected as a LLM node's context in a chatbot

create a new Knowledge and import an empty document

• document name: whatever

• Chunk settings: Automatic

• Index mode: Economical

1. download LongTermMemory.yml file and import it using Dify "Import DSL file"

2. In the imported workflow, config MODEL for LLM:Parameter Extractor node, which LLM performance must be >= gpt4. Tips: gpt4 series, deepseek-chat are recommended, gpt3.5-turbo was tested but not powerful enough to run this tool.

add the "LongTermMemory" tool before LLM node

notes for parameters:

• base_url: http://host.docker.internal for docker-compose deployment on MacOS(use http://docker.for.win.localhost for Windows 10/11), append the port like :8082 if not using default nginx port config

• dataset_id: dataset id

• document_id: document id

• api_key: Dify Knowledge API key, this is not "Chat App API key". You can get one on 'Knowledge -> API ACCESS' page, a Knowledge API key should be look like "dataset-xxx"

• memory_template: for example

<UserInfo>
  <name>${user_name}</name>
  <age>${user_age}</age>
  <hobbies>${user_hobbies}</hobbies>
  <any_user_info>${any_user_info}</any_user_info>
</UserInfo>

• user_input: {{#sys.query#}}

CONTEXT：select LongTermMemory node's text

SYSTEM prompt:

{{#context#}}

user input:

• who am I?

  • chatbot should reply no idea

• I'm Rain

  • check out the knowledge page, there should be a new segment with user_id, name, and keyword attribute tags

• I like coding

  • check out the knowledge page, the segment should be updated, tag should be updated

• I like reading, also

  • check out the knowledge page, the segment should be updated, tag should be updated

Debug and Preview -> Restart (make sure to start a new conversion without chat history)

user input:

• what's my name?
  • chatbot should reply current user's name
• what do I like?
  • chatbot should reply current user's hobbies

Run App (You can use different browsers to access and impersonate different users)

user input:

• I'm Jerry
• I like eating
  • check out the knowledge page, there should be a new segment with different user_id, name, and keyword attribute tags